Hi Mike,
That seems very odd that small files work better than large files. First try
reading this, I wrote it and it might help:
http://www.samag.com/documents/s=1146/sam0109j/0109j.htm , OR,
http://www.sysadminmag.com/articles/2001/0109/0109a/0109a.htm
Try these:
-Since you have two RAID arrays, or hdisks, I would use the LVM to stripe the
LV that your storage pool is on, over the two arrays. Make sure that the LV
that your storage pool is on, over the two arrays. Make sure that is it not
'spread out', but create a striped LV: mklv -S 128k, for example
Here is my LV:
root@unxr:/dev>lslv dbdumps_lv
LOGICAL VOLUME: dbdumps_lv VOLUME GROUP: ADSMvg128
LV IDENTIFIER: 00008107f034d2ed.2 PERMISSION: read/write
VG STATE: active/complete LV STATE: opened/syncd
TYPE: jfs WRITE VERIFY: off
MAX LPs: 2170 PP SIZE: 128 megabyte(s)
COPIES: 1 SCHED POLICY: striped
LPs: 2170 PPs: 2170
STALE PPs: 0 BB POLICY: non-relocatable
INTER-POLICY: maximum RELOCATABLE: no
INTRA-POLICY: edge UPPER BOUND: 2
MOUNT POINT: /home/dbdumps LABEL: /home/dbdumps
MIRROR WRITE CONSISTENCY: on
EACH LP COPY ON A SEPARATE PV ?: yes
STRIPE WIDTH: 2
The two disks, or 'STRIPE WIDTH', are RAID 0 arrays.
-Make sure that the queue depth on your adapter is set properly, for each array:
array:
root@unxr:/dev>lsattr -El hdisk6
pvid 00008107347137090000000000000000 Physical volume identifier
False
*** queue_depth 64 *** Queue depth
True
write_queue_mod 3 Write queue depth modifier
True
adapter_a ssa0 Adapter connection
False
adapter_b none Adapter connection
False
primary_adapter adapter_a Primary adapter
True
reserve_lock yes RESERVE device on open
True
connwhere_shad 81073AF833FB4CE SSA Connection Location
False
max_coalesce 0x20000 Maximum coalesced operation
True
size_in_mb 145781 Size in Megabytes
False
location Location Label
True
-Try and put each array on different loop, or better a different adapter
make sure that the VMM is tuned. Check the minfree and minfree free make sure
that the VMM is tuned. Check the minfree and minfree free settings and
especially the read ahead, as well as others. The command is:
/usr/samples/kernel/vmtune.
This is my command for my TSM server, with a very similar setup to yours:
/usr/samples/kernel/vmtune -f 256 -F 768 -u 32 -p 20 -P 70 -r 2 -R 64 -b 128 -l
65536 -B 547 -w 32000
Here is a speed test:
I copied a 2 GB file to /dev/null in one window. Ran iostat 10 in another:
tty: tin tout avg-cpu: % user % sys % idle % iowait
1.6 515.1 47.5 26.0 16.7 9.8
Disks: % tm_act Kbps tps Kb_read Kb_wrtn
hdisk3 8.5 49.6 10.7 240 256
hdisk1 2.7 15.6 3.0 16 140
hdisk2 0.9 6.0 1.2 0 60
hdisk0 2.1 10.8 2.7 0 108
**** hdisk6 67.8 19407.8 153.5 194256 16 ****
hdisk5 0.0 0.0 0.0 0 0
hdisk4 0.0 0.4 0.1 0 4
**** hdisk7 80.3 19240.0 151.4 192588 4 ****
I have a look at the article though, it gives lots of good ideas, I hope!
Miles
-------------------------------------------------------------------------------------------
----------------
Miles Purdy
Miles Purdy
System Manager
Farm Income Programs Directorate
Winnipeg, MB, CA
purdym AT fipd.gc DOT ca
ph: (204) 984-1602 fax: (204) 983-7557
-------------------------------------------------------------------------------------------
----------------
>>> Michael.MA.Wiggan AT PDO.CO DOT OM 05-Nov-01 1:20:15 AM >>>
>>> Michael.MA.Wiggan AT PDO.CO DOT OM 05-Nov-01 1:20:15 AM >>>
Hi Guru's
I have recently installed a new TSM 4.1.3 server on a P600 and AIX 4.1.3. I
have 2*75Gybyte RAID5 Disk Cache attached through SSA. The performance for
reading from disk to tape is fine for small files, but I only get 2.5MB/Sec
when files are larger than say 1Gbyte. I have set "SELFTUNETXNsize Yes",
and it brought it up tothis speed when using two drives.
I have 18 *4Gbyte files on each RAID5 and can see that the data is evenly
distributed in each file. I can see that the file say 7Gbytes is spread
throughout many of the Disk Storage Volumes. I don't want to go straight to
tape as this causes a huge headache in scheduling.
Is this a known bottleneck or is there a way around to get at least
20MB/Sec?
Kind Regards
Mike Wiggan, TCS/31
Infrastructure Integration Specialist
Petroleum Devlopment Oman LLC
(michael.ma.wiggan AT pdo.co DOT om)
|